基于混合专家模型的岩石薄片图像分类  

Classification of Rock Thin Section ImagesBased on Mixture of Expert Model

作　　者：周程阳 刘伟[2] 吴天润 李骜 韩霄松[1] ZHOU Chengyang;LIU Wei;WU Tianrun;LI Ao;HAN Xiaosong(College of Software,Jilin University,Changchun 130012,China;CNPC Engineering Technology R&D Company Limited,Beijing 102206,China)

机构地区：[1]吉林大学软件学院,长春130012 [2]中国石油集团工程技术研究院,北京102206

出　　处：《吉林大学学报（理学版）》2024年第4期905-914,共10页Journal of Jilin University:Science Edition

基　　金：国家重大科技专项基金(批准号:2011ZX05044,2011ZX05001);国家自然科学基金(批准号:62372494);吉林省科技发展计划项目(批准号:20220201145GX,20220601112FG);大学生创新创业训练计划项目(批准号:202310183221)。

摘　　要：以常见的5种岩石薄片作为研究对象构建数据集,提出一种新的基于混合专家模型的岩石薄片图像分类模型.该模型从薄片图像中学习到每种岩石图像的特征,并对其进行分类.首先,使用多个基于卷积神经网络(CNN)和Transformer的图像分类模型(ResNet50,MobileNetV3,InceptionV3,DeiT等)对数据进行训练;其次,选取效果较好的模型,通过构建混合专家模型,得到最终的预测结果,其岩性识别准确率(ACC)和AUC在验证集上达到85.33%和96.69%,在测试集上达到87.16%和96.75%;最后,通过混合专家模型结合多个模型,综合各模型的优势,平衡各模型间的贡献,提高分类结果的准确性和鲁棒性,使得到的分类结果更可靠、稳定.We proposed a new classification of rock thin section images based on mixture of expert model by using five common rock thin sections as the research object to construct a dataset.The model learned the characteristics of each rock image from the thin section images and classified them.Firstly,multiple image classification models based on convolutional neural network(CNN)and Transformer(such as ResNet50,MobileNetV3,InceptionV3,DeiT,etc.)were used to train the data.Secondly,models with better performance were selected,a mixture of experts model was built to obtain the final prediction result.The ACC and AUC of lithology recognition reached 85.33%and 96.69%on the validation set and 87.16% and 96.75% on the test set.Finally,by combining a mixture of experts model with multiple models,combining advantage of each model,balancing their contributions between each model,we improved the accuracy and robustness of classification results,making the obtained classification results more reliable and stable.

关 键 词：岩石薄片分类 混合专家模型 图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]